Qwen3-Embedding-4B

Qwen3-Embedding-4B, published on 5 June 2025, sits between the two sizes people usually choose from. It turns a document into a vector of up to 2,560 numbers, works over a 32,768-token window and covers more than 100 languages. Against the 8B flagship it gives up 1.13 points of multilingual MTEB (69.45 against 70.58) while halving the memory a server has to hold, which for a large index is the difference between one graphics card and two. One result runs the other way, and it is the reason this profile exists separately. On instruction-following retrieval — the task where a search request arrives as a written sentence rather than a keyword, measured by the Inst. Retri. column of MTEB multilingual — the 4B scores 11.56 against the 8B's 10.06. It is the best figure in Alibaba's whole comparison table, more than twice Google's gemini-embedding-exp at 5.18. Bigger is not uniformly better even inside one family trained the same way, and this is the single column where the pattern breaks. The architecture follows Qwen3-4B-Base: 36 layers, hidden dimension 2,560, 32 attention heads over 8 key-value heads. Output vectors shorten on demand to as few as 32 numbers, so an index can be compressed after the fact. The model accepts a task instruction prepended to the input, which the vendor values at 1 to 5 percent accuracy and recommends writing in English. One limit worth stating for readers building a Chinese-language index: at 72.27 on C-MTEB this model still finishes below ritrieve_zh_v1, a Chinese specialist of 0.3 billion parameters scoring 72.71. Thirteen times the size does not buy an advantage in a single language. Weights are Apache 2.0, with a separate GGUF build. Adoption in the 30 days to 1 September 2026: 3.22 million Hugging Face downloads and 315 likes.

▸Videos

No videos yet.

▸Reviews